Tightening up methane plume source rate estimation in EnMAP and PRISMA images

Abstract. Reducing methane emissions from human activities is essential to tackle climate change. To monitor these emissions, we rely on satellite observations, which enable regular, global-scale tracking. Methane emissions are typically quantified by their source rate – the mass of gas emitted per unit of time. Our goal here is to estimate the emission source rate of methane plumes detected by hyperspectral imagers such as PRISMA or EnMAP. For this task, we generated a large synthetic dataset using large eddy simulation (LES) to train a deep learning model. This dataset was specifically designed to avoid network overfitting with careful plume temporal sampling and plume scaling. Our deep learning network, MetFluxNet, does not require any wind information or a plume mask. Moreover, it accurately predicts the source rate even in the presence of false positives. MetFluxNet performs well on our dataset with a mean absolute percentage error (MAPE) of 8.3 % over a wide range of source rates from 500 to 25 000 kg h−1. Notably, it remains effective at lower source rates, where background noise is typically high. To validate its real-world applicability, we tested MetFluxNet on real plumes with known ground-truth fluxes. The predicted source rates fell systematically within the 95 % confidence intervals, demonstrating its reliability for real-world plume estimation. Finally, in a comparison with recent state-of-the-art methods, MetFluxNet outperformed the deep learning-based S2MetNet and the physics-based integrated mass enhancement (IME) method.